The task of conveying the coffee grain’s quality characteristics from cultivation and post-harvest shaping falls to the coffee roasting operation. The machine operator must adjust the equipment’s working parameters in real time in addition to knowing the intrinsic qualities of the bean in order to produce the best expression of a coffee or a particular profile for a target market. Decisions then frequently depend on the operator’s experience and are subjective. The objective prediction of the evolution of pertinent variables during coffee roasting is presented in this work using mathematical model-based methodologies. The temperature distribution of a coffee mass being roasted in rotary drum equipment was predicted using a theoretical technique coupling a one-way CFD-DEM model. The validated model was used to simulate scenarios with air recirculation, obtaining improvements in roast homogeneity and energy use. A data-driven machine learning (ML) strategy using Support Vector Machines (SVM) and Artificial Neural Networks (ANN) was applied to assess the influence of operative conditions on the coffee roasting process’s efficiency. Thereby, a multilevel factorial design was set up for each machine learning technique, considering different hyperparameters to optimize the best architecture with which to describe the roasting process. The results revealed the potential of SVM and ANN as reliable tools to describe the roasting process of coffee. Results are promising for inline implementation of coffee roasting process monitoring in a multivariate approach.